On Mon, 31 Jan 2011, Kevin Hilman wrote:

For the on-chip SoC devices we're managing with OMAP, we're currently
only using one set: post ops on [runtime_]suspend and pre ops on
[runtime_]resume.

However, I could imagine (at least conceptually) using the pre ops on
suspend to do some constraints checking and/or possibly some
management/notification of dependent devices.  Another possiblity
(although possibly racy) would be using the pre ops on suspend to
initiate some high-latency operations.

Dependency management is very relevant here, since we're talking about
relations that explicitly aren't of the parent-child type.  If any of
the devices in question get marked for async suspend/resume, for
example, they certainly will need dependency handling.

I guess the main problem with two sets is wasted space.  e.g, if I move
OMAP to this (already hacking on it) there will be only 2 functions used
in post ops: [runtime_]suspend() and 2 used in pre ops [runtime_]_resume().

The wasted space is minimal; we're only talking about one extra pm_ops
structure for each power domain.  Presumably any reasonable SoC isn't
going to have a tremendous number of separate power domains.  Or am I
wrong about this?

Alan Stern